If you run voice agents in production, you already know the pattern: the stack performs great in a controlled demo, then meets the real world and starts failing calls in ways nobody can quite explain.
Most teams reach for transcript review or manual call sampling to work out what went wrong, but those tools look downstream, after the damage is done. They rarely catch the actual culprit, which is often the audio itself.
That's the gap Berlin-based ai-coustics is going after with Tyto, an Audio Insight model it launched today.
Tyto scores every call in real time and predicts whether the incoming audio conditions are going to cause failures further down the pipeline, in components like VAD, ASR, and speech-to-speech models. When it flags a problem, it tells you why, breaking the signal down across seven acoustic dimensions instead of leaving you to guess. The point is to surface audio quality as a root cause right at the audio layer, before it ever reaches a model.
Two ways teams can put it to work:
Post-call analysis: see which calls failed because of audio, and what specifically drove it.
Real-time adaptive behavior: let your agent stack respond to changing acoustic conditions mid-call.
"Voice AI has a blind spot," said Fabian Seipel, co-founder and CEO of ai-coustics. "Builders invest heavily in their models and prompts, but have no systematic way to know whether audio quality is degrading performance in the field. Tyto closes that gap."
Teams already running ai-coustics in production echo that. "With ai-coustics on, we've had the best customer satisfaction scores of all the tests we've done," said Razvan Kusztos, VP of Engineering at PolyAI. Seb Hapte-Selassie, co-founder at telli, was blunter: "At our scale, audio quality isn't a detail. It's the difference between converting a customer and losing one. ai-coustics gives us the clarity our agents need to perform at volume."
ai-coustics says its tech is trusted in production by teams including PolyAI, telli, LiveKit, Phonely, and TrusstAI.
Tyto is available now through the ai-coustics SDK at ai-coustics.com.
Based on information provided by ai-coustics.
